Quick answerPrice a multi-year AI agent renewal against what the work is worth today, not against the original contract's per-unit pricing carried forward, and structure the term to include a scheduled mid-contract repricing checkpoint so a multi-year commitment does not lock either side into economics that stop reflecting reality within the first year.
This assumes the decision to renew is already made
Evaluating whether to renew an AI vendor at all is a separate decision that should already be settled before this question comes up. This post starts from the other side of that decision: you are renewing, the term is multi-year, and the model landscape has moved enough since signing that the original pricing no longer reflects the actual cost or value of the work.
Why old pricing does not carry forward cleanly
A contract signed two or three years ago was priced against the cost and capability of models available at the time. Since then, inference costs for comparable capability have typically fallen, newer models often do more per call, and what it costs to run an AI agent at scale has shifted in ways that were not knowable when the original terms were set. Carrying forward a per-conversation or per-seat price that was calibrated to old economics either overpays the vendor for capability that is now commodity, or, if the vendor tries to hold the customer to old volume assumptions, underprices what the vendor is now actually delivering.
The renewal negotiation should start from a fresh baseline: what would this scope of work cost to build and run today, at today's model economics, rather than anchoring on the prior contract's numbers and negotiating a percentage change from there.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Structuring the term to avoid the same problem next time
A multi-year commitment made sense when model capability moved more slowly. It is a riskier bet now, so the contract structure itself should account for that, not just the initial price. Two structural elements do most of the work: a shorter initial commitment with renewal options rather than a single long lock-in, and a scheduled repricing checkpoint, for example at the 12 or 18 month mark, tied to an objective reference such as published inference pricing for comparable capability, rather than left to a full renegotiation from scratch.
This is a different mechanism from renegotiating an AI vendor contract during a budget cut, which is triggered by the customer's financial pressure. A scheduled repricing checkpoint is planned into the contract from the start, triggered by the passage of time and model-landscape change rather than by either side's financial distress, and is easier for both parties to agree to precisely because it is not adversarial.
What to lock in versus what to leave flexible
Not everything should reprice automatically. Service levels, support terms, and data handling commitments should generally stay fixed for the full term, since negotiating SLAs and support terms once per contract is worth the stability. What should flex is the unit economics tied directly to model cost: per-call or per-token pricing, volume tiers, and any minimum commitment sized against a specific model's capability. Separating these two categories in the contract language up front avoids a renewal conversation two years from now trying to renegotiate everything at once.
FAQ
Should the company always push for a shorter contract now?
Not automatically. A shorter term reduces the risk of stale pricing but can cost leverage on price and priority support. The repricing checkpoint inside a longer term is often a better trade than a much shorter commitment.
What reference point should a repricing checkpoint use?
An objective, publicly checkable one, such as published API pricing for comparable model capability at the checkpoint date, so neither side has to rely on the other's internal cost claims.
Does this apply to internally built agents too?
The pricing mechanics are vendor-specific, but the underlying discipline, repricing against current model economics rather than carrying forward stale assumptions, applies just as much to internal budget planning for an internally built agent.